Short-term load forecasting method based on weather classification

By combining a weather-based classification method with long-sequence and point-by-point load forecasting models, and utilizing meteorological and historical data for power load forecasting, the problem of insufficient accuracy in existing power load forecasting technologies is solved, and high-precision power load forecasting is achieved.

CN123121504APending Publication Date: 2026-11-17STATE GRID JIBEI ELECTRIC POWER COMPANY +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611313262.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies struggle to achieve high-precision power load forecasting, especially after renewable energy is integrated into the power system, which exacerbates the randomness and nonlinearity of power load, leading to insufficient grid stability and load forecasting accuracy.

Method used

By using a weather classification-based approach, meteorological data and historical load data for the forecast period are obtained, the target weather type is determined, and forecasts are made using long-series load forecasting models and point-by-point load forecasting models, respectively. The forecast accuracy is improved through data fusion and correction.

Benefits of technology

It improves the accuracy and precision of power load forecasting, reduces interference from irrelevant features, enhances the stability and adaptability of forecasts, and is applicable to power load forecasting under different weather conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123121504A_ABST
    Figure CN123121504A_ABST
Patent Text Reader

Abstract

This application provides a short-term load forecasting method based on weather classification. The method includes: acquiring meteorological and date data for the period to be forecasted, as well as first historical load data for the preceding period; determining the target weather type for the period to be forecasted based on the meteorological data; selecting correlation feature data corresponding to the target weather type from the meteorological data, date data, and first historical load data to obtain correlation feature data; inputting the correlation feature data into a preset long-sequence load forecasting model and a point-by-point load forecasting model corresponding to the target weather type, respectively, to obtain first load forecast data output by the long-sequence load forecasting model and second load forecast data output by the point-by-point load forecasting model; and determining the final load forecast data for the period to be forecasted based on the first and second load forecast data. This application can improve the accuracy of power load forecasting and achieve high-precision load forecasting.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of power data technology, and in particular to a short-term load forecasting method based on weather classification. Background Technology

[0002] With the integration of a large amount of renewable energy into the power system, electricity load has become increasingly complex and volatile, exacerbating the instability of energy consumption and posing a significant challenge to the stable operation of the power grid. Accurate electricity load forecasting is crucial for effectively dispatching electricity, peak shaving and valley filling, strengthening the dynamic balance between supply and demand on the generation and consumption sides, and improving energy utilization and system flexibility.

[0003] In related technologies, power load forecasting mainly employs methods such as mathematical statistics and machine learning, including support vector regression and artificial neural networks. However, with the increasing randomness and nonlinearity of power load data, these methods struggle to achieve high-precision load forecasting, resulting in low accuracy in power load forecasting. Summary of the Invention

[0004] This application provides a short-term load forecasting method based on weather classification to improve the accuracy of power load forecasting and achieve high-precision load forecasting.

[0005] In a first aspect, embodiments of this application provide a short-term load forecasting method based on weather classification, including: Obtain meteorological and date data for the period to be predicted, as well as the first historical load data for the period preceding the period to be predicted; Based on the meteorological data, determine the target weather type for the forecast period; The associated feature data is obtained by selecting the associated feature data corresponding to the target weather type from the meteorological data, the date data, and the first historical load data; The associated feature data are respectively input into the long-sequence load forecasting model and the point-by-point load forecasting model corresponding to the preset target weather type to obtain the first load forecasting data output by the long-sequence load forecasting model and the second load forecasting data output by the point-by-point load forecasting model. Based on the first load forecast data and the second load forecast data, the final load forecast data for the period to be forecasted is determined.

[0006] Secondly, embodiments of this application provide a short-term load forecasting device based on weather classification, comprising: The acquisition module is used to acquire meteorological data and date data for the period to be predicted, as well as the first historical load data of the period preceding the period to be predicted; The classification module is used to determine the target weather type for the forecast period based on the meteorological data. The filtering module is used to select data with associated features corresponding to the target weather type from the meteorological data, the date data, and the first historical load data to obtain associated feature data; The prediction module is used to input the associated feature data into a long-sequence load prediction model and a point-by-point load prediction model corresponding to a preset target weather type, respectively, to obtain the first load prediction data output by the long-sequence load prediction model and the second load prediction data output by the point-by-point load prediction model. The correction module is used to determine the final load forecast data for the period to be predicted based on the first load forecast data and the second load forecast data.

[0007] Thirdly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect or any possible implementation of the first aspect.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in the first aspect or any possible implementation thereof.

[0009] Fifthly, embodiments of this application provide a computer program product that, when run on an electronic device, causes the electronic device to perform the steps of the method described in the first aspect or any possible implementation thereof.

[0010] The beneficial effects of the embodiments in this application compared with the prior art are: This application embodiment determines the target weather type for the forecast period using meteorological data of the forecast period, enabling targeted forecasting based on the target weather type. It selects relevant feature data corresponding to the target weather type from the meteorological date data and the first historical load data, obtaining related feature data. This allows for the filtering of highly correlated features for power load forecasting under the target weather type, reducing interference from low-correlation features and ensuring the accuracy of power load forecasting. Load forecasting is performed using both a long-sequence load forecasting model and a point-by-point load forecasting model, yielding long-sequence first load forecast data and point-by-point second load forecast data. Load forecasting can be performed using both methods. Finally, the final load forecast data is determined using the first and second load forecast data, allowing for point-by-point correction of the forecast data and further improving the accuracy of power load forecasting. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the implementation of the short-term load forecasting method based on weather classification provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of the TiDE model provided in the embodiments of this application; Figure 3 This is a schematic diagram of the model corresponding to the short-term load forecasting method based on weather classification provided in the embodiments of this application; Figure 4 This is a schematic diagram of the clusters corresponding to sunny days provided in the embodiments of this application; Figure 5 This is a schematic diagram of the clusters corresponding to high-temperature days provided in the embodiments of this application. Figure 6 This is a schematic diagram of the clusters corresponding to rainy and snowy days provided in the embodiments of this application; Figure 7 This is a heatmap showing the correlation between features and predicted power load provided in the embodiments of this application; Figure 8 These are schematic diagrams illustrating the prediction results of the various models provided in the embodiments of this application on a sunny day; Figure 9 These are schematic diagrams illustrating the prediction results of various models provided in the embodiments of this application on hot days. Figure 10These are schematic diagrams illustrating the prediction results of various models provided in the embodiments of this application on rainy and snowy days; Figure 11 This is a schematic diagram showing the results of analyzing the hidden layer size provided in an embodiment of this application; Figure 12 This is a schematic diagram showing the results of analyzing the number of encoder layers provided in an embodiment of this application; Figure 13 This is a schematic diagram illustrating the results of analyzing the number of decoder layers provided in an embodiment of this application; Figure 14 This is a schematic diagram of the structure of the short-term load forecasting device based on weather classification provided in the embodiments of this application; Figure 15 This is a schematic diagram of the electronic device provided in the embodiments of this application. Detailed Implementation

[0013] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0014] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0015] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0016] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0017] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0018] Furthermore, the term "multiple" mentioned in the embodiments of this application should be interpreted as two or more.

[0019] The inventors have discovered that methods such as mathematical statistics and machine learning are commonly used for power load forecasting. However, with the increasing randomness and nonlinearity of power load data, achieving high-precision load forecasting becomes difficult. Therefore, it is necessary to consider a new method for power load forecasting.

[0020] In order to improve the accuracy of power load forecasting, in the embodiments of this application, the target weather type for the forecast period is determined by meteorological data of the forecast period, and targeted forecasting is performed based on the target weather type; the correlation features corresponding to the target weather type are selected, and load forecasting is performed using a long-sequence load forecasting model and a point-by-point load forecasting model respectively, and point-by-point correction is performed to obtain the final load forecasting data, thereby improving the forecasting accuracy of power load forecasting.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following description will be provided in conjunction with the accompanying drawings and specific embodiments.

[0022] Figure 1 The implementation flowchart of the weather classification-based short-term load forecasting method provided in the embodiments of this application is described in detail below: Step 101: Obtain meteorological data and date data for the period to be predicted, as well as the first historical load data of the period preceding the period to be predicted.

[0023] In this embodiment, the duration of the time period to be predicted and the duration of the preceding time period can be the same. For example, the duration of the time period to be predicted is 24 hours, and the duration of the preceding time period is also 24 hours. Alternatively, the durations of the two time periods can be different, and the preceding time period can be the hour preceding the time period to be predicted.

[0024] Here, the period preceding the predicted period refers to the period immediately preceding the predicted period. For example, if the predicted period is September 22nd, then the preceding period could be September 21st. Alternatively, if the predicted period is from 9:00 AM to 12:00 PM (including the time point corresponding to 9:00 AM), the preceding period could be from 6:00 AM to 9:00 AM (excluding the time point corresponding to 9:00 AM), or it could be from 8:00 AM to 9:00 AM (excluding the time point corresponding to 9:00 AM).

[0025] The first historical load data of the period preceding the period to be predicted is the load data collected in the previous period. Load data is collected at 15-minute intervals. If the previous period is the day before the period to be predicted, the first historical load data can be the load data corresponding to the 96 collection points of the previous day. If the previous period is 4 hours before the period to be predicted, the first historical load data can be the load data corresponding to the 16 collection points of the period preceding the period to be predicted.

[0026] Meteorological data can include temperature, humidity, rainfall, and snowfall, while date data can include time period, date, holidays, and non-holidays.

[0027] Step 102: Based on meteorological data, determine the target weather type for the forecast period.

[0028] In this embodiment, considering that power load data changes in different ways depending on the weather, different weather types can be classified according to meteorological data. When making a forecast, the target weather type corresponding to the forecast period is found based on the meteorological data of the period to be forecasted.

[0029] Here, weather types can include sunny days, hot days, and rainy / snowy days.

[0030] Optionally, in this embodiment, the target weather type for the forecast period can be determined based on meteorological data by first obtaining typical meteorological parameters of multiple preset weather types; then calculating the first distance between the forecast period and each weather type based on the meteorological data and typical meteorological parameters; and finally, determining the weather type corresponding to the smallest first distance as the target weather type for the forecast period.

[0031] In this embodiment, by calculating the first distance between the time period to be predicted and each weather type, the similarity between the weather conditions of the time period to be predicted and each weather type can be obtained. The smaller the first distance, the more similar the weather conditions of the time period to be predicted are to that weather type. The weather type corresponding to the smallest first distance is the most similar weather type, so the time period to be predicted can be classified into the range of that weather type, thereby determining the weather type of the time period to be predicted.

[0032] Step 103: Select the data of the correlation features corresponding to the target weather type from the meteorological data, date data and first historical load data to obtain the correlation feature data.

[0033] In this embodiment, the factors or features with high correlation differ when predicting power load data for different weather types. Therefore, corresponding features can be selected for prediction based on different weather types.

[0034] Here, the correlation features corresponding to the target weather type are those features that have a high correlation with the predicted power load data under that weather type.

[0035] Step 104: Input the associated feature data into the long-sequence load forecasting model and the point-by-point load forecasting model corresponding to the preset target weather type, respectively, to obtain the first load forecasting data output by the long-sequence load forecasting model and the second load forecasting data output by the point-by-point load forecasting model.

[0036] In this embodiment, a long-sequence load forecasting model and a point-by-point load forecasting model are used for load forecasting. The long-sequence load forecasting model can directly obtain load forecasting data for a sequence, while the point-by-point load forecasting model can perform load forecasting for each time point, thereby obtaining load forecasting data for each time point within the forecast period.

[0037] Here, each weather type corresponds to a long-sequence load forecasting model and a point-by-point load forecasting model. When forecasting, the long-sequence load forecasting model and the point-by-point load forecasting model corresponding to the target weather type of the period to be forecasted are selected for load forecasting in order to make accurate forecasts.

[0038] Step 105: Determine the final load forecast data for the period to be forecasted based on the first load forecast data and the second load forecast data.

[0039] In this embodiment, the first load forecast data can be corrected point by point using the second load forecast data to improve the forecast accuracy.

[0040] Optionally, the period to be predicted includes multiple prediction points. In this embodiment, the final load prediction data for the period to be predicted is determined based on the first load prediction data and the second load prediction data. This can be achieved by: calculating the average of the first load prediction data and the second load prediction data corresponding to each prediction point in the period to be predicted, respectively, to obtain the mean data for each prediction point in the period to be predicted; and then obtaining the final load prediction data for the period to be predicted based on the mean data.

[0041] In this embodiment, the first load forecast data and the second load forecast data can be averaged and fused to combine the forecast data obtained from the two forecasting methods, so as to accurately predict the final load forecast data for the period to be predicted.

[0042] This application embodiment determines the target weather type for the forecast period using meteorological data of the forecast period, enabling targeted forecasting based on the target weather type. It selects relevant feature data corresponding to the target weather type from the meteorological date data and the first historical load data, obtaining related feature data. This allows for the filtering of highly correlated features for power load forecasting under the target weather type, reducing interference from low-correlation features and ensuring the accuracy of power load forecasting. Load forecasting is performed using both a long-sequence load forecasting model and a point-by-point load forecasting model, yielding long-sequence first load forecast data and point-by-point second load forecast data. Load forecasting can be performed using both methods. Finally, the final load forecast data is determined using the first and second load forecast data, allowing for point-by-point correction of the forecast data and further improving the accuracy of power load forecasting.

[0043] In some embodiments, before obtaining typical meteorological parameters for various preset weather types, the method may further include: Step 1: Obtain historical meteorological data for multiple historical periods preceding the period to be predicted.

[0044] Step 2: Randomly select a preset number of historical time periods as cluster centers.

[0045] Here, the preset quantity can be 3, 4, or 5, etc., depending on the area to be predicted.

[0046] It should be noted that in this embodiment, the actual historical time period is selected as the cluster center when selecting the cluster center and when iterating, so as to use real data for clustering and improve the ability to resist outliers.

[0047] Step 3: Based on the historical meteorological data for each historical period and the historical meteorological data corresponding to the cluster center, calculate the second distance from each historical period to each cluster center.

[0048] Step 4: Assign each historical period to the cluster containing its corresponding target cluster center; where the target cluster center for a historical period is the cluster center corresponding to the smallest second distance of that historical period.

[0049] Step 5: Calculate the total distance from each historical time period within each cluster to other historical time periods within that cluster, and take the historical time period with the smallest total distance within each cluster as the new cluster center; complete one iteration of cluster center calculation.

[0050] In this embodiment, the historical period with the smallest total distance within each cluster is used as the new cluster center. The actual historical period is used as the cluster center, and the selection is made from the existing historical periods. This reduces the sensitivity to outliers and can increase the robustness to noise and outliers.

[0051] Step 6: Iterate continuously until the cluster center remains unchanged or the maximum number of iterations is reached. Based on the cluster center that remains unchanged or the maximum number of iterations is reached, determine a variety of preset weather types, and use the historical meteorological data of the historical period corresponding to the cluster center as the typical meteorological parameters of the weather type corresponding to that cluster center.

[0052] In this embodiment, clustering can be considered complete when the cluster centers no longer change or when the preset maximum number of iterations is reached.

[0053] Optionally, considering potential anomalies or data loss during the acquisition of power load and meteorological data, this paper first preprocesses the raw data to improve the accuracy of the prediction. In the preprocessing step, we use linear interpolation to fill in missing values ​​in the data. This method is simple and effective, and the calculation formula is as follows: In the formula, Indicates missing values. The x-coordinate represents the missing values. The x-coordinate to the left of the missing value. This represents the value to the left of the missing value. The x-coordinate to the right of the missing value. This represents the value to the right of the missing value.

[0054] In some embodiments, before selecting the associated feature data corresponding to the target weather type from meteorological data, date data, and first historical load data to obtain the associated feature data, the following may also be done: Step 1: Divide multiple historical time periods according to various preset weather types to obtain the historical time periods corresponding to each weather type.

[0055] In the above embodiments, historical meteorological data were clustered to classify each historical time period and obtain the corresponding weather type for each historical time period. In this embodiment, the historical time periods corresponding to each weather type are obtained by classifying the data according to the weather type corresponding to each historical time period, thereby facilitating model training and prediction for each weather type.

[0056] Step 2: Obtain historical meteorological data, historical date data, second historical load data, and third historical load data for each historical period corresponding to each weather type.

[0057] In this embodiment, each weather type is used as a category to further divide the data of all historical time periods, thereby obtaining the data corresponding to each weather type, so as to train and predict the model for each weather type.

[0058] Here, the second historical load data is the load data corresponding to each historical time period, and the third historical load data is the load data corresponding to the previous time period of each historical time period.

[0059] If each historical period is a continuous historical period, such as one, two, or three years before the period to be predicted, then the third historical load data and the second historical load data have overlapping data. In this case, the historical load data of all historical periods and the period before all historical periods can be directly obtained as the total historical load data. The historical load data corresponding to each historical period can be extracted from it as the second historical load data, and the historical load data of the period before each historical period can be extracted from it as the third historical load data.

[0060] Step 3: For each weather type, train a short-term load forecasting model for that weather type based on the corresponding historical meteorological data, historical date data, third historical load data, and second historical load data.

[0061] Step four: During the training process of the short-term load forecasting model for this weather type, determine the frequency of use and gain of each feature in historical meteorological data, historical date data, and third historical load data.

[0062] Step 5: Determine the characteristic importance of each characteristic for short-term load forecasting of this weather type based on the frequency of use and gain of each characteristic.

[0063] Features whose importance is greater than a preset value are identified as the associated features corresponding to this weather type.

[0064] In this embodiment, a short-term load forecasting model can be trained for each weather type to identify the features with high relevance when forecasting power load under each weather type.

[0065] Here, machine learning models can be used for load prediction training, such as the eXtremeGradient Boosting (XGBoost) model and the random forest model.

[0066] Optionally, the XGBoost model can be used. For each weather type, the corresponding historical meteorological data, historical date data, third historical load data, and second historical load data are input into the XGBoost feature correlation analysis module. This module can be used to comprehensively analyze the correlation between features, identify and extract key features, i.e., the associated features corresponding to that weather type. This can effectively eliminate features with low correlation, minimize their interference with the prediction process, and ensure the accuracy of the model.

[0067] XGBoost is an efficient and robust gradient boosting decision tree algorithm that excels at managing large datasets and complex high-dimensional features. As an ensemble learning technique built on gradient boosting, XGBoost improves model performance by sequentially building and refining a series of decision trees. Each new tree progressively improves the model by reducing the residuals (prediction errors) of the previous tree.

[0068] The specific steps are as follows: Step 1: Make initial predictions for all training samples using a simple model (such as a constant model); Step 2: Calculate the residual for each sample, which is the difference between the actual value and the predicted value. The formula is: ,in, Let i be the residual of the i-th sample. This is the actual value. These are predicted values.

[0069] Step 3: Use the residuals as the target value to fit a new decision tree. The predictions from this tree are used to update the model's predictions. The predictions from the new decision tree are scaled by the learning rate to prevent overfitting.

[0070] The corresponding formula is: ,in, For the predicted values ​​of the new decision tree The predicted values ​​are from the old decision tree. For learning rate, For the new tree to the sample The predicted value.

[0071] Step 4: Update the model's predictions. Repeat steps 2 and 3 until the predetermined number of trees is reached or the model performance no longer shows significant improvement.

[0072] The objective function and regularization conditions are as follows: The objective function contains a loss function and a regularization term, which are used to balance the model's fit and complexity.

[0073] The objective function is ,in, For loss functions (such as mean squared error, cross-entropy). For regularization terms, Let k be the k-th tree.

[0074] Regularization terms are used to control the complexity of the model and prevent overfitting. Their formula can be: ;in, Control leaf nodes Quantity, Controlling leaf node weights L2 regularization.

[0075] In some embodiments, before inputting the associated feature data into a long-series load forecasting model and a point-by-point load forecasting model corresponding to a preset target weather type, respectively, and obtaining the first load forecasting data output by the long-series load forecasting model and the second load forecasting data output by the point-by-point load forecasting model, the method further includes: For each weather type, select the data of the associated features corresponding to the weather type from the historical meteorological data, historical date data and third historical load data of the historical period corresponding to the weather type to obtain the historical feature data of each historical period corresponding to the weather type; Based on the historical characteristic data and second historical load data corresponding to this weather type, the preset time series prediction model is trained to obtain the long-series load prediction model corresponding to this weather type. Based on the historical characteristic data and second historical load data corresponding to this weather type, the preset point prediction model is trained to obtain the point-by-point load prediction model corresponding to this weather type.

[0076] In this embodiment, a time series prediction model and a point prediction model can be trained for each weather type to make predictions from both the time series and point perspectives, thereby improving the accuracy of the predictions.

[0077] Here, the point prediction model can adopt the XGBoost model in the above embodiments, and the time series prediction model can adopt the Temporal Information-Driven Encoder-Decoder (TiDE) model. The TiDE model is an encoder-decoder architecture that utilizes a multilayer perceptron (MLP) framework, specifically designed to simplify long-term time series forecasting. This model cleverly combines the simplicity and fast execution of linear models while effectively managing covariates and nonlinear dependencies. Under certain conditions, a simplified linear version of the TiDE model can achieve near-optimal error rates for linear dynamic systems (LDS), demonstrating its significant potential in terms of high prediction accuracy and computational efficiency. Its structure is as follows... Figure 2 As shown.

[0078] (1) Feature projection steps The feature projection step is a crucial component of the TiDE model, primarily focusing on dynamic covariates—input variables that change over time, such as weather conditions or holidays. In this step, the model compresses these covariates from a high-dimensional space to a low-dimensional space, simplifying the processing of the required information. This reduction not only lowers computational complexity but also enhances the model's ability to extract valuable features from the covariates, thereby improving the accuracy of time series predictions. Essentially, feature projection compresses the input data, allowing for more efficient processing by the encoder and decoder.

[0079] (2) Dense Multilayer Perceptron (MLP) Encoder In the TiDE model, the dense multilayer perceptron (MLP) encoder plays a crucial role. Its main task is to transform historical time-series data (the retrieval part) and relevant covariates affecting prediction (such as external factors like dates and weather) into internal representations. This transformation is performed through a series of MLP network layers, each progressively transforming and learning from the input data to capture complex patterns and dependencies in the time series. Essentially, the encoder compresses the raw input into a condensed form of information that the model can then further analyze and process.

[0080] (3) Dense MLP decoder In the TiDE model, the dense MLP decoder is crucial for transforming encoded information back into time series predictions. Once it receives the internal representation from the encoder, the decoder begins constructing predictions for future time points. The decoder employs specially trained multilayer perceptron layers to identify patterns in the encoded data and maps these patterns to future covariates, thereby generating predicted values ​​for the time series. Essentially, the decoder's role is to interpret the encoded data and transform it into accurate predictions of future events.

[0081] (4) Timing Decoder The temporal decoder is a fundamental component of the TiDE model, specifically designed to process the output of the dense decoder. It operates at each time step, transforming the decoder's predictions into the final time-series output. By applying a tailored transformation at each time point, the temporal decoder refines the temporal dependencies of the predictions, thereby enhancing the model's ability to grasp the temporal dynamics of time-series data. Essentially, the temporal decoder integrates the decoding process with the time dimension, producing accurate step-by-step predictions that significantly improve the accuracy and applicability of the forecasts.

[0082] (5) Global residual connection Global linear residual connections are commonly implemented in deep neural networks to improve learning efficiency and simplify training. In the TiDE model, residual connections help transfer information directly from earlier layers to deeper structures. This helps maintain the integrity of the original input data throughout the learning process and effectively alleviates the vanishing gradient problem common in deep networks. Essentially, residual connections act like a shortcut, allowing input data to bypass multiple intermediate layers and directly influence the final output, thus simplifying the flow of information within the model.

[0083] Optionally, in this embodiment, to effectively train the model and ensure accurate validation, the dataset can be divided into a training set, a validation set, and a test set in a 7:2:1 ratio. The data is arranged chronologically, with the first 21 days of each month serving as the training set, the middle 6 days as the validation set, and the remaining days as the test set. This arrangement captures patterns observed throughout the year, improving training effectiveness. To eliminate dimensionality differences in the data, accelerate model convergence, and improve training accuracy, the dataset is normalized. The normalization formula is as follows: In the formula, This represents the normalized data value. This represents the minimum value of the sample data. This represents the maximum value of the sample data. This represents the data value before normalization.

[0084] Correspondingly, after obtaining the forecast results, they can be denormalized to obtain the final load forecast data.

[0085] Optionally, before training the preset time series prediction model based on the historical characteristic data and second historical load data corresponding to the weather type to obtain the long-series load prediction model corresponding to the weather type, the hyperparameters of the time series prediction model can be optimized by grid search method.

[0086] Accordingly, before training the preset point prediction model based on the historical characteristic data and second historical load data corresponding to the weather type to obtain the point-by-point load prediction model corresponding to the weather type, the hyperparameters of the point prediction model can be optimized by the grid search method.

[0087] In some specific embodiments, mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) are selected as indicators, and their definitions are as follows: ; ; ; In the formula, Indicates the actual value. This represents the measured value, and N is the number of data samples.

[0088] To verify the effectiveness of the prediction methods in the above embodiments, actual data from September 2021 to September 2022 was used, including 35,040 detailed records from a specific region, with each record sampled at 15-minute intervals. This dataset includes power load data for the region, as well as relevant meteorological parameters such as temperature, air pressure, rainfall, snowfall, wind speed, and wind direction.

[0089] The prediction is performed using the method proposed in the embodiments of this application. The prediction model can be found in [reference needed]. Figure 3 As shown, it comprises three main stages: First, a clustering algorithm is used to cluster the load data into three typical load types: sunny, hot, and rainy / snowy weather. Next, XGBoost association analysis is performed on these data types to identify the most influential features. Subsequently, models are developed and trained for these three load types. Finally, the performance of the model in this embodiment is compared with other state-of-the-art models such as LSTM, Transformer, and Autoformer. A detailed analysis of the load forecasting accuracy under various climatic conditions is provided to demonstrate the superiority and practicality of the model proposed in this application.

[0090] (1) Clustering results All electricity load data were classified using meteorological factors such as load size, temperature, humidity, rainfall, and snowfall as clustering criteria. The results categorized the load data into three main types: sunny days, hot days, and rainy / snowy days. Figures 4 to 6 As shown, in each category, the gray curve represents all load curves in that category, and the red curve represents the cluster center curve for each category. The graph shows that the load values ​​for sunny days are generally lower, the load values ​​for hot days are slightly higher with similar fluctuations to the sunny days, while the load values ​​for rainy / snowy days are also higher, but with more pronounced fluctuations.

[0091] (2) Correlation analysis Based on extensive experience and in-depth data analysis, this paper selects historical load data, meteorological data, and calendar rule data as input features for the model. Load data includes the load sequence one hour prior to the forecast time to capture short-term load trends; meteorological data includes key meteorological indicators such as temperature, humidity, rainfall, and snowfall to assess the potential impact of meteorological factors on load changes. Given the periodicity of load data within a day, week, and between holidays and non-holidays, this study also performs single-time encoding of calendar rules for hours, weeks, and holidays.

[0092] The XGBoost algorithm is used to analyze the correlation between features of different categories. During training, XGBoost evaluates the contribution of each feature to the decision-making process and prioritizes the features with the highest benefit. In this way, the model can naturally identify predictive features that have a significant impact on the outcome. The correlation heatmap is shown below. Figure 7 As shown.

[0093] from Figure 7 As can be seen, these three load categories are highly correlated with historical load data, demonstrating a high sensitivity to changes in historical loads. Furthermore, the correlation with historical load data tends to be stronger as the forecast target time approaches. Regarding weather data, all three load categories are highly sensitive to temperature, with the correlation between temperature and load being most significant on hot days. In contrast, humidity and wind speed show lower correlations across all three load categories. The rain and snow load category shows strong correlations with both rainfall and snowfall. Additionally, analysis of calendar data indicates that the three load categories are more sensitive to holidays, while weekday and hour coding show lower correlations and feature contributions. Based on the results of the feature correlation analysis, the most relevant features were selected, as shown in Table 1.

[0094] Table 1 Feature Selection Table (3) Model parameter settings To further verify the effectiveness and superiority of the proposed model, it is compared with several benchmark models, including LSTM, Transformer and Autoformer models.

[0095] The LSTM model, with its unique structure, solves the problem of traditional RNNs in capturing long-term dependencies and can effectively model time series data. In contrast, the Transformer and Autoformer models significantly improve performance by introducing attention mechanisms, especially when dealing with long-term data, achieving excellent results on sequence data.

[0096] To ensure the objectivity and fairness of the experiment, the model's hyperparameters were adjusted using grid search and optimized using validation set loss. Furthermore, the hyperparameters of other benchmark models were also adjusted to ensure they achieved optimal predictive performance, thus reflecting the fairness of the experiment. The final model's hyperparameter results are shown in Table 2.

[0097] Table 2 Model Parameter Table (4) Multi-model comparison experiment To comprehensively evaluate the performance of the model presented in this paper, typical days representing different climatic conditions were selected to compare prediction results. Specifically, typical days for each weather category were predicted, including sunny, hot, rainy, and snowy days. Each prediction contained 96 time-series data points, and the model's input step size was also set to 96 points. The prediction results are shown in the figure below. Figure 8 , Figure 9 and Figure 10 It is evident that the model demonstrates excellent predictive accuracy under all weather conditions and closely matches the actual data. This fully illustrates the model's outstanding performance in handling complex and ever-changing weather conditions.

[0098] Table 3 shows the accuracy of each model's predictions across different categories. Clearly, the proposed model achieves high accuracy in predicting both static and fluctuating loads, highlighting its stability and reliability. These results confirm that the model effectively captures the periodic characteristics of static sequences and accurately represents the complex patterns of fluctuating sequences. Compared to other prediction models, the model in this application exhibits the smallest error value and the highest fitting coefficient, further demonstrating its superior performance in handling dynamic and fluctuating load data.

[0099] When evaluating the performance of the four models, the model proposed in this application demonstrated superior performance under all weather conditions, with lower mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE) compared to the other models. For example, on sunny days, the model's MAE was 8.49, RMSE was 10.29, and MAPE was 2.55%, significantly outperforming Autoformer (MAE 18.29, RMSE 22.33, MAPE 5.50%), Transformer (MAE 21.34, RMSE 27.59, MAPE 6.40%), and LSTM (MAE 25.97, RMSE 30.60, MAPE 7.80%). Among these models, the Autoformer model generally performed second best, consistently outperforming Transformer and LSTM. Notably, under high-temperature conditions, the Transformer's MAE and RMSE were 25.51 and 31.95, respectively, lower than LSTM's 28.66 and 34.81. Although LSTM has the weakest performance, its prediction results still largely conform to the load trend.

[0100] Table 3 Comparison of Prediction Accuracy The proposed model demonstrates significant performance advantages under various weather conditions. For example, under clear weather conditions, its MAE, RMSE, and MAPE are 53.6%, 54.0%, and 53.7% lower than the closest Autoformer model, respectively. Under high-temperature conditions, these metrics are 41.9%, 42.4%, and 42.0% lower, respectively. Under rainy or snowy conditions, the model's performance advantage is even more pronounced, with MAE, RMSE, and MAPE values ​​being 55.9%, 59.2%, and 55.9% lower than Autoformer, respectively. In contrast, traditional LSTM and Transformer models struggle under complex or extreme weather conditions. These substantial performance differences highlight the model's superior accuracy and robustness in managing forecasting tasks across diverse weather conditions.

[0101] (5) Comparison of model training and inference time Furthermore, the training and inference times of each baseline model were compared against the proposed model. A rigorous runtime comparison is as follows: The training times for the models are as follows: LSTM model 355 min, Transformer model 300 min, Autoformer model 280 min, and the model proposed in this application 80 min.

[0102] The prediction times for the models are as follows: 31s for the LSTM model, 24s for the Transformer model, 12s for the Autoformer model, and 8s for the model proposed in this application.

[0103] During the training phase, this model achieved the best training efficiency among all benchmark models. During the test inference phase, the model in this application was significantly faster than other benchmark models. This is primarily because LSTM processes sequences stepwise, while Transformer and Autoformer models require additional time to compute attention.

[0104] (6) Ablation test To investigate the impact of each component on the overall predictive performance of the model, this application conducted a series of ablation experiments under the same experimental environment, dataset, and model parameter settings. All experiments were based on a 96-step prediction step size and included the following model variants: the original TiDE model, the clustering + TiDE model, the TiDE + XGBoost pointwise correction model, and the integrated model proposed in this paper. The prediction accuracy of each variant is detailed in Table 4.

[0105] Table 4 Comparison of ablation experiments As shown in Table 4, the model proposed in this application exhibits the best performance under all weather conditions, with the lowest MAE, RMSE, and MAPE values ​​regardless of whether the weather is sunny, hot, or rainy / snowy. For example, under sunny conditions, the model's MAE is 8.49, RMSE is 10.29, and MAPE is 2.55%, significantly outperforming other models. In contrast, the XGBoost+TiDE model has an MAE of 11.46, RMSE of 16.77, and MAPE of 5.13% under sunny conditions, but it still outperforms the standalone TiDE model (MAE 17.95, RMSE 23.46, MAPE 6.06%) and the k-Medoids+TiDE model (MAE 12.10, RMSE 15.46, MAPE 4.04%), highlighting the effectiveness of pointwise calibration. Under high-temperature conditions, the model proposed in this application continues to lead, with a MAE of 9.12, RMSE of 10.95, and MAPE of 2.62%, while the XGBoost+TiDE model outperforms other variants (MAE of 12.45, RMSE of 15.54, and MAPE of 3.01%). Under rain and snow conditions, this model maintains the lowest MAE, RMSE, and MAPE values, at 7.17%, 8.91%, and 2.07%, respectively. The XGBoost+TiDE model's MAPE of 2.94% also surpasses the k-Medoids+TiDE model's 3.51%.

[0106] Overall, the model described in this application consistently demonstrates the highest accuracy and stability under various weather conditions, with the XGBoost+TiDE configuration generally outperforming both the k-Medoids+TiDE and standalone TiDE models. The results of these ablation experiments conclusively demonstrate the effectiveness of each component, thus affirming the comprehensive superiority of this approach.

[0107] (7) Sensitivity analysis of model parameters like Figure 11 , Figure 12 and Figure 13 As shown, this application also conducted a sensitivity analysis on three parameters: hidden layer size, number of encoder layers, and number of decoder layers. The range of hidden layer size is [128, 256, 512, 1024]. Larger hidden layers may require more training data and time for effective training, but may also lead to overfitting. When set to 1024, the model will overfit, resulting in increased error and excessively long training time. The range of numEncoderLayers and numDecoderLayers is [1, 2, 3, 4]. When these values ​​are too low, the model will have difficulty capturing more complex patterns and relationships, leading to increased error. Conversely, when the values ​​are too high, training time and error will also increase significantly.

[0108] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0109] The following are device embodiments of this application. For details not described in detail, please refer to the corresponding method embodiments described above.

[0110] Figure 14 A schematic diagram of the structure of a short-term load forecasting device based on weather classification provided in an embodiment of this application is shown. For ease of explanation, only the parts related to the embodiment of this application are shown, and are described in detail below: like Figure 14 As shown, the short-term load forecasting device 140 based on weather classification includes: The acquisition module 1401 is used to acquire meteorological data and date data for the period to be predicted, as well as the first historical load data of the period preceding the period to be predicted; Classification module 1402 is used to determine the target weather type for the forecast period based on meteorological data; The filtering module 1403 is used to select data with associated features corresponding to the target weather type from meteorological data, date data and first historical load data to obtain associated feature data; The prediction module 1404 is used to input the associated feature data into the long-sequence load prediction model and the point-by-point load prediction model corresponding to the preset target weather type, respectively, to obtain the first load prediction data output by the long-sequence load prediction model and the second load prediction data output by the point-by-point load prediction model. The correction module 1405 is used to determine the final load forecast data for the period to be forecasted based on the first load forecast data and the second load forecast data.

[0111] In one possible implementation, the classification module 1402 is specifically used for: Obtain typical meteorological parameters for various preset weather types; Based on meteorological data and typical meteorological parameters, calculate the first distance between the forecast period and each weather type; The weather type corresponding to the smallest first distance is determined as the target weather type for the period to be predicted.

[0112] In one possible implementation, the time period to be predicted includes multiple prediction points; The correction module 1405 is specifically used for: Calculate the average of the first load forecast data and the second load forecast data corresponding to each forecast point in the forecast period, and obtain the mean data of each forecast point in the forecast period. Based on the average data, the final load forecast data for the period to be predicted is obtained.

[0113] In one possible implementation, the weather-classification-based short-term load forecasting device 140 further includes a training module, which is specifically used for: Obtain historical meteorological data for multiple historical periods preceding the period to be predicted; A preset number of historical time periods are randomly selected as cluster centers; Based on the historical meteorological data for each historical period and the historical meteorological data corresponding to the cluster centers, the second distance from each historical period to each cluster center is calculated. Each historical period is assigned to the cluster containing its corresponding target cluster center; where the target cluster center for a historical period is the cluster center corresponding to the smallest second distance of that historical period. Calculate the total distance from each historical time period within each cluster to other historical time periods within that cluster, and take the historical time period with the smallest total distance within each cluster as the new cluster center; complete one iteration of cluster center calculation; The process iterates continuously until the cluster center remains unchanged or the maximum number of iterations is reached. Based on the cluster center that remains unchanged or the maximum number of iterations is reached, multiple preset weather types are determined, and the historical meteorological data of the historical period corresponding to the cluster center is used as the typical meteorological parameters of the weather type corresponding to that cluster center.

[0114] In one possible implementation, the training module is also used for: Based on multiple preset weather types, several historical time periods are divided to obtain the historical time periods corresponding to each weather type; Acquire historical meteorological data, historical date data, second historical load data, and third historical load data for each historical period corresponding to each weather type; For each weather type, a short-term load forecasting model is trained based on the historical meteorological data, historical date data, third historical load data, and second historical load data corresponding to that weather type. During the training process of the short-term load forecasting model for this weather type, determine the frequency of use and gain of each feature in historical meteorological data, historical date data, and third-party historical load data; Based on the frequency of use and gain of each feature, determine the feature importance of each feature in the short-term load forecast for this weather type; Features whose importance is greater than a preset value are identified as the associated features corresponding to this weather type.

[0115] In one possible implementation, the training module is also used for: For each weather type, select the data of the associated features corresponding to the weather type from the historical meteorological data, historical date data and third historical load data of the historical period corresponding to the weather type to obtain the historical feature data of each historical period corresponding to the weather type; Based on the historical characteristic data and second historical load data corresponding to this weather type, the preset time series prediction model is trained to obtain the long-series load prediction model corresponding to this weather type. Based on the historical characteristic data and second historical load data corresponding to this weather type, the preset point prediction model is trained to obtain the point-by-point load prediction model corresponding to this weather type.

[0116] Figure 15 This is a schematic diagram of an electronic device provided in an embodiment of this application. Figure 15 As shown, the electronic device 150 of this embodiment includes: a processor 151, a memory 152, and a computer program 153 stored in the memory 152 and executable on the processor 151. When the processor 151 executes the computer program 153, it implements the steps in the various embodiments of the weather-based short-term load forecasting method described above, for example... Figure 1 Steps 101 to 105 are shown. Alternatively, when processor 51 executes computer program 153, it implements the functions of each module in the above-described device embodiments, for example... Figure 14 The functions of modules 141 to 143 are shown.

[0117] For example, computer program 153 can be divided into one or more modules / units, one or more of which are stored in memory 152 and executed by processor 151 to complete this application. One or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 153 in electronic device 150. For example, computer program 153 can be divided into... Figure 14 Modules 141 to 143 are shown.

[0118] Electronic device 150 may include, but is not limited to, processor 151 and memory 152. Those skilled in the art will understand that... Figure 15 This is merely an example of electronic device 150 and does not constitute a limitation on electronic device 150. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0119] The processor 151 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0120] The memory 152 can be an internal storage unit of the electronic device 150, such as a hard disk or RAM of the electronic device 150. The memory 152 can also be an external storage device of the electronic device 150, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, FlashCard, etc., equipped on the electronic device 150. Furthermore, the memory 152 can include both internal and external storage units of the electronic device 150. The memory 152 is used to store computer programs and other programs and data required by the electronic device. The memory 152 can also be used to temporarily store data that has been output or will be output.

[0121] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0122] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0123] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0124] In the embodiments provided in this application, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0125] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0127] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0128] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A short-term load forecasting method based on weather classification, characterized in that, include: Obtain meteorological and date data for the period to be predicted, as well as the first historical load data for the period preceding the period to be predicted; Based on the meteorological data, determine the target weather type for the forecast period; The associated feature data is obtained by selecting the associated feature data corresponding to the target weather type from the meteorological data, the date data, and the first historical load data; The associated feature data are respectively input into the long-sequence load forecasting model and the point-by-point load forecasting model corresponding to the preset target weather type to obtain the first load forecasting data output by the long-sequence load forecasting model and the second load forecasting data output by the point-by-point load forecasting model. Based on the first load forecast data and the second load forecast data, the final load forecast data for the period to be forecasted is determined.

2. The short-term load forecasting method based on weather classification according to claim 1, characterized in that, The step of determining the target weather type for the forecast period based on the meteorological data includes: Obtain typical meteorological parameters for various preset weather types; Based on the meteorological data and the typical meteorological parameters, calculate the first distance between the forecast period and each weather type; The weather type corresponding to the smallest first distance is determined as the target weather type for the period to be predicted.

3. The short-term load forecasting method based on weather classification according to claim 1, characterized in that, The time period to be predicted includes multiple prediction points; Based on the first load forecast data and the second load forecast data, the final load forecast data for the period to be forecasted is determined, including: Calculate the average of the first load forecast data and the second load forecast data corresponding to each forecast point in the forecast period to obtain the mean data of each forecast point in the forecast period. Based on the mean data, the final load forecast data for the period to be predicted is obtained.

4. The short-term load forecasting method based on weather classification according to claim 2, characterized in that, Before obtaining typical meteorological parameters for various preset weather types, the process also includes: Obtain historical meteorological data for multiple historical periods preceding the period to be predicted; A preset number of historical time periods are randomly selected as cluster centers; Based on the historical meteorological data for each historical period and the historical meteorological data corresponding to the cluster center, calculate the second distance from each historical period to each cluster center; Each historical period is assigned to the cluster containing its corresponding target cluster center; where the target cluster center for a historical period is the cluster center corresponding to the smallest second distance of that historical period. Calculate the total distance from each historical time period within each cluster to other historical time periods within that cluster, and take the historical time period with the smallest total distance within each cluster as the new cluster center; complete one iteration of cluster center calculation; The process iterates continuously until the cluster center remains unchanged or the maximum number of iterations is reached. Based on the cluster center that remains unchanged or the maximum number of iterations is reached, multiple preset weather types are determined, and the historical meteorological data of the historical period corresponding to the cluster center is used as the typical meteorological parameters of the weather type corresponding to that cluster center.

5. The short-term load forecasting method based on weather classification according to claim 4, characterized in that, Before obtaining the associated feature data by selecting the associated feature data corresponding to the target weather type from the meteorological data, the date data, and the first historical load data, the method further includes: Based on multiple preset weather types, the multiple historical time periods are divided to obtain the historical time periods corresponding to each weather type; Acquire historical meteorological data, historical date data, second historical load data, and third historical load data for each historical period corresponding to each weather type; For each weather type, a short-term load forecasting model is trained based on the historical meteorological data, historical date data, third historical load data, and second historical load data corresponding to that weather type. During the training process of the short-term load forecasting model for this weather type, determine the frequency of use and gain of each feature in historical meteorological data, historical date data, and third-party historical load data; Based on the frequency of use and gain of each feature, determine the feature importance of each feature in the short-term load forecast for this weather type; Features whose importance is greater than a preset value are identified as the associated features corresponding to the weather type.

6. The short-term load forecasting method based on weather classification according to claim 5, characterized in that, Before inputting the associated feature data into the preset long-series load forecasting model and point-by-point load forecasting model corresponding to the target weather type, respectively, to obtain the first load forecasting data output by the long-series load forecasting model and the second load forecasting data output by the point-by-point load forecasting model, the method further includes: For each weather type, select the data of the associated features corresponding to the weather type from the historical meteorological data, historical date data and third historical load data of the historical period corresponding to the weather type to obtain the historical feature data of each historical period corresponding to the weather type; Based on the historical characteristic data and second historical load data corresponding to this weather type, the preset time series prediction model is trained to obtain the long-series load prediction model corresponding to this weather type. Based on the historical characteristic data and second historical load data corresponding to this weather type, the preset point prediction model is trained to obtain the point-by-point load prediction model corresponding to this weather type.

7. The short-term load forecasting method based on weather classification according to claim 6, characterized in that, Before training a pre-defined time-series forecasting model based on historical characteristic data and second historical load data corresponding to the weather type to obtain a long-series load forecasting model corresponding to the weather type, the following steps are also included: The hyperparameters of the time series prediction model are optimized using a grid search method. Before training a pre-defined point prediction model based on historical characteristic data and second historical load data corresponding to the weather type to obtain a point-by-point load prediction model corresponding to the weather type, the following steps are also included: The hyperparameters of the point prediction model are optimized using a grid search method.

8. A short-term load forecasting device based on weather classification, characterized in that, include: The acquisition module is used to acquire meteorological data and date data for the period to be predicted, as well as the first historical load data of the period preceding the period to be predicted; The classification module is used to determine the target weather type for the forecast period based on the meteorological data. The filtering module is used to select data with associated features corresponding to the target weather type from the meteorological data, the date data, and the first historical load data to obtain associated feature data; The prediction module is used to input the associated feature data into a long-sequence load prediction model and a point-by-point load prediction model corresponding to a preset target weather type, respectively, to obtain the first load prediction data output by the long-sequence load prediction model and the second load prediction data output by the point-by-point load prediction model. The correction module is used to determine the final load forecast data for the period to be predicted based on the first load forecast data and the second load forecast data.

9. An electronic device comprising a memory and a processor, the memory for storing a computer program, the processor for calling and running the computer program stored in the memory, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.